tim-shu/drug-affinity-rf-baseline

Drug affinity: frozen Random Forest baseline

Predict pKi from a complete small-molecule SMILES for one of 162 human protein targets. Higher pKi means tighter binding: pKi = −log10(Ki in mol/L). Select the protein by UniProt accession using targets.json.

The repository contains one trained RandomForestRegressor per target, from the frozen RF_full_descriptors, seed 42. Each forest has 100 trees, min_samples_leaf=1, max_features="sqrt". Inputs are chiral Morgan fingerprints (radius 2, 2048 bits) and eight properties: molecular weight, logP, TPSA, HBD, HBA, rotatable bonds, rings and heavy atoms. All disconnected components, charges and specified stereochemistry are retained. Protein IDs select a model; they are not numerical features.

Run

Download the model repository (about 362 MB of forests):

from huggingface_hub import snapshot_download
snapshot_download("tim-shu/drug-affinity-rf-baseline", local_dir="rf_baseline")

Install huggingface_hub for the download, then use Python 3.12 and the supplied inference dependencies:

python -m pip install -r rf_baseline/requirements.txt
python rf_baseline/affinity_predictor.py --target P00918 --smiles 'NS(=O)(=O)c1ccccc1'

Or in Python, after adding the downloaded folder to the import path:

import sys
sys.path.insert(0, "rf_baseline")
from affinity_predictor import AffinityBaseline
baseline = AffinityBaseline("rf_baseline")
pki = baseline.predict(["NS(=O)(=O)c1ccccc1"], target_id="P00918")
print(pki)

Weights use joblib serialization; use the exact scikit-learn and RDKit versions in requirements.txt. prediction_examples.json contains reference predictions for all supported targets, including multicomponent structures. verification.json records the local inference check; manifest.json lists SHA-256 checksums.

Data and evaluation

Training data and executed report, configuration baseline_benchmark: 204,847 pairs, with 163,330 train, 20,427 validation and 21,090 test pairs. Sources: ChEMBL, BindingDB and DrugCentral. Full-structure/protein pairs define labels; common chemical-scaffold groups define partitions. The weights are trained on train only, without a train+validation refit.

Selection used target-macro validation MAE averaged over seeds 42/43/44. The selection score was 0.594075 pKi. The uploaded seed-42 model has:

Split	Macro MAE	Macro RMSE	Macro Spearman
Validation	0.594005	0.775525	0.691222
Test	0.575935	0.747746	0.693083

Metrics weight protein targets equally. Some underlying data were examined in earlier project experiments, so test is not a new blind evaluation. final_baseline.json and training_manifest.json record exact selection, input hashes and training settings; reproducible training code is in the dataset repository's baseline_study/reproduction/ folder.

Scope

This is a research baseline for molecular affinity. It supports the listed proteins and does not generalize directly to unseen proteins. Accuracy can degrade for structures unlike training compounds; experimental conditions, pH and binding poses are not explicit inputs. Predicted affinity does not establish a therapeutic effect or a new clinical indication.

Source-data attribution and licenses are documented in the linked dataset repository.